Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,31 @@
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
---
|
| 4 |
+
|
| 5 |
+
# T5-ANCE
|
| 6 |
+
|
| 7 |
+
T5-ANCE generally follows the training procedure described in [this page](dr-msmarco-passage.md), but uses a much larger batch size.
|
| 8 |
+
|
| 9 |
+
Dataset used for training:
|
| 10 |
+
- MS MARCO Passage
|
| 11 |
+
|
| 12 |
+
Evaluation result:
|
| 13 |
+
|
| 14 |
+
|Dataset|Metric|Result|
|
| 15 |
+
|---|---|---|
|
| 16 |
+
|MS MARCO Passage (dev) | MRR@10 | 0.3570|
|
| 17 |
+
|
| 18 |
+
Important hyper-parameters:
|
| 19 |
+
|
| 20 |
+
|Name|Value|
|
| 21 |
+
|---|---|
|
| 22 |
+
|Global batch size|256|
|
| 23 |
+
|Learning rate|5e-6|
|
| 24 |
+
|Maximum length of query|32|
|
| 25 |
+
|Maximum length of document|128|
|
| 26 |
+
|Template for query|`<text>`|
|
| 27 |
+
|Template for document|`Title: <title> Text: <text>`|
|
| 28 |
+
|
| 29 |
+
### Paper
|
| 30 |
+
|
| 31 |
+
\-
|